Interlock trains long-horizon agents on typed completion graphs

What the research found

Interlock synthesizes long-horizon agent tasks as typed completion graphs and fine-tunes Qwen3.6 on accepted trajectories, improving AutomationBench, JobBench and Harvey LAB pass rates by about 3 to 6 points.

Claims and how well the evidence supports them

1. Interlock training improves AutomationBench task pass rates by 4.00-4.50 percentage points 99%
2. Interlock training improves JobBench weighted scores by 3.20-6.64 points 99%
3. Interlock fine-tunes two Qwen3.6 backbones on their own accepted trajectories 99%

Evidence

• huggingface.co https://huggingface.co/papers/2610.19686

Actionability rated low.